Labels

🌊 Mapping Hidden Flood Impacts With AI: How

 Satellite Embeddings Reveal Geomorphic Change

Floods leave scars on the landscape, some obvious, some subtle, and many invisibles to the human eye. Traditional flood mapping identifies surface water during an event, but it often fails to capture underlying geomorphic changes: sediment rearrangements, vegetation stress, channel migration, and long‑term floodplain response.

In my latest research at the University of Alabama, I developed a new approach using Google’s Alpha Earth Satellite Embeddings and a geomorphic metric called the Geomorphic Anomaly Index (GAI) to detect these subtle flood‑driven changes over time.

This blog introduces the method, shows two key figures from the analysis, and explains why this approach opens new doors for post‑flood impact assessment.

πŸ“Œ What is the Geomorphic Anomaly Index (GAI)?

Most flood-mapping techniques focus on water, but water is temporary. The GAI instead measures changes in the land surface itself by comparing high‑dimensional satellite embeddings (64‑band vectors) between consecutive years.

These embeddings encode:

  • texture
  • spectral signatures
  • vegetation condition
  • soil exposure
  • morphology
  • surface roughness

and more.

By analyzing the distance between embeddings from year to year, we can quantify how much the landscape changed—then statistically determine whether those changes were unusual.

That “unusualness” is expressed as a Z‑score, which becomes the GAI.

  • GAI ≈ 0 → normal year‑to‑year variability
  • GAI > 2 → unusually large geomorphic disturbance
  • GAI < 0 → quieter‑than‑normal year

πŸ“ˆ Figure 1 — GAI Over Time Reveals the 2018–

2019 Flood Signature

Below is the figure you provided, showing the GAI trajectory across seven consecutive year-pairs for a 3‑km river buffer around the Waccamaw River.


Figure 1. Geomorphic Anomaly Index (GAI) across consecutive year pairs.
The 2018–2019 pair (red) corresponds to the major flooding associated with Hurricane Florence and is excluded from baseline calculations.


πŸ” What the
Figure Shows

  1. 2018–2019 stands out
    The GAI spikes sharply, indicating geomorphic disturbance far above typical year‑to‑year variability. This aligns with NOAA and USGS documentation of Hurricane Florence flooding, which produced widespread inundation and geomorphic disruption.

  2. 2020–2021 shows secondary disturbance
    Likely associated with high precipitation years and elevated discharge.

  3. 2022–2023 shows a negative anomaly
    A quieter geomorphic year—possibly vegetation recovery, reduced flood intensity, or a sediment‑stabilizing period.

  4. Excluding the main flood year from the baseline
    ensures the anomaly calculation remains stable and unbiased.

This plot demonstrates that AI‑generated embeddings can detect geomorphic impacts consistent with hydrologic records, even without needing direct surface water inputs.

πŸ—Ί️ Figure 2 — Hydrologic Context: USGS Stations,

 Tributaries, and Basin Structure


Figure 2. Map of the Waccamaw River basin showing USGS stations, HUC12 pour points, the main channel (red), and tributary network (blue).

πŸ” What the Figure Shows

  • Main stem of the Waccamaw River (red)
    A sinuous system that drains toward coastal South Carolina, sensitive to backwater and rainfall-driven flooding.

  • Dense tributary network (light blue)
    Provides numerous hydrologic pathways where geomorphic changes can accumulate.

  • USGS stations (blue stars & orange triangles)
    Offer streamflow validation points up to 20 km from the main stem.

  • HUC12 pour points (black dots)
    Illustrate sub-watershed boundaries, useful for scaling GAI analysis.

This hydrologic network provides the backbone for validating the AI‑derived anomaly signals.

⭐ Why Satellite Embeddings Are a Gamechanger

Traditional flood mapping is built on water detection. But floods:

  • kill vegetation
  • deposit sediment
  • scour banks
  • restructure floodplains
  • reshape channels

Most satellite algorithms miss these changes because they look exclusively for water signatures.

Satellite embeddings, on the other hand:

  • compress multi-sensor information into a stable, machine-learned representation
  • are robust to cloud noise
  • integrate multiple spectral, textural, and physical cues
  • work even when water is no longer visible
  • measure geomorphic disturbance, not just water extent

In your research, this allowed you to detect:

  • the signature of Hurricane Florence, even one year after the event
  • secondary disturbances associated with regional flooding
  • quiet recovery years where the landscape stabilized

This is powerful for emergency management, watershed planning, and climate‑risk assessment.

πŸ§ͺ Methods in Plain Language

Here’s the workflow simplified into four steps:

1. Compute annual embeddings (2017–2024)

Each year produces a 64‑dimensional feature vector per pixel.

2. Compute Ξ”E: the “distance” between consecutive years

This measures how much the landscape changed between Year A and Year B.

3. Build a baseline

Use all “normal” years to estimate typical variability.

4. Compute GAI

A Z-score that measures how unusual the target year-pair is relative to baseline.

If GAI > 2 → significant geomorphic disturbance.

πŸŒͺ️ What the Data Reveal: Floods Leave Long-

Lasting Geomorphic Signatures

The 2018–2019 spike corresponds to Florence, which delivered:

  • record-setting rainfall
  • extreme river discharge
  • prolonged inundation
  • widespread sediment redistribution

Even though water disappears from satellite images after a few days, the landscape remembers the flood.

The GAI captures that memory.

🏞️ How This Helps Real-World Flood Analysis

πŸ“ 1. Detect “invisible floods”

Some floods don’t appear in surface water maps due to clouds, timing, or sensor limitations.

πŸ“ 2. Quantify recovery time

GAI over time shows whether a basin takes:

  • 1 year to recover
  • 2 years
  • or several years (important for ecosystem monitoring)

πŸ“ 3. Improve flood hazard models

GAI can be used as training data for machine‑learning flood forecasting.

πŸ“ 4. Support FEMA or state agencies

Embedding-derived anomalies offer evidence of flood impact even when field data are sparse.

🌍 A New Era of Post‑Flood Intelligence

The combination of:

  • high-dimensional satellite embeddings
  • year-to-year geomorphic change detection
  • statistical anomaly scoring
  • hydrologic context from USGS stations

creates a new way to understand flood impacts beyond water extent alone.

Your research shows that AI can extract rich geomorphic signals across an entire watershed—signals that traditional remote sensing techniques have historically struggled to capture.

This has applications in:

  • climate resilience planning
  • flood risk modeling
  • watershed rehabilitation
  • river restoration
  • FEMA flood assessments
  • insurance and infrastructure planning
Labels:

Comments

Post a Comment